Visual Question Answering (VQA) integrates computer vision and natural language processing, enabling models to respond to questions related to images. Existing approaches often rely on simplistic attention mechanisms, resulting in insufficient comprehension of complex visual features, which limits their reasoning capabilities. Additionally, traditional Vision Transformer (ViT) exhibit shortcomings in leveraging visual information effectively. To address these challenges, this paper presents the KV-KBVQA model, which integrates Kan-ViT with Cross-Modal Feature Fusion Based Bidirectional Residual Attention, specifically designed to process visual information and optimize the interaction between visual and textual modalities. Evaluation results demonstrate that the KV-KBVQA model achieves an accuracy of 54.55% on the challenging OK-VQA dataset, representing an improvement of 0.77% over the baseline, thereby highlighting its potential in advancing VQA tasks.